OpenVLA-7B — LIBERO-Spatial LoRA (20k steps, with gradient clipping)
Fine-tuned from openvla/openvla-7b on the libero_spatial_no_noops task suite using LoRA (rank=32) for 20,000 gradient steps across two sequential runs on an NVIDIA GH200 (ARM64/aarch64).
Key contribution: includes a gradient clipping fix missing from the upstream OpenVLA training script — without it, training collapses at ~6k steps due to an unclamped gradient spike.
Evaluation Results
Evaluated on LIBERO-Spatial (10 tasks, 5 trials per task = 50 episodes total).
Setup
Steps
Score
Official paper (x86, flash-attn)
10k
84.7%
Official openvla-7b-finetuned-libero-spatial (ARM64 eval)
10k
37/50 = 74%
Our run 1 — no gradient clipping
10k
20/50 = 39%
Our run 2 — with gradient clipping
10k
28/50 = 56%
This checkpoint — grad clipping, 20k steps (ARM64 eval)
20k
36/50 = 72%
Our 20k checkpoint matches the official published 10k checkpoint (74% vs 72%, within statistical noise at 5 trials/task) on the same ARM64 hardware, confirming the gradient clipping fix is sufficient to recover training stability.
The ~12pp gap vs. the paper (84.7%) is a hardware effect: the official paper used x86 with native flash attention; on ARM64 the pre-built flash-attn binary is x86-only. Flash-attn 2.5.8 was built from source for run 3 but not run 2, so training used eager attention for most of the run. The official checkpoint evaluated at 74% on the same ARM64 setup, confirming this gap is hardware-driven, not a model quality issue.
Green = Run 2 (steps 0–10k, starting from base openvla/openvla-7b)
Blue = Run 3 (steps 10k–20k, starting from the run 2 merged checkpoint)
Loss descends steadily from ~2.4 → ~0.68. Action token accuracy rises from ~0.05 → ~0.82. No loss spikes — contrast with run 1 (no gradient clipping) which spiked at step ~6k and never recovered.
Training Details
Setting
Value
Base model
openvla/openvla-7b
Dataset
libero_spatial_no_noops
Total steps
20,000 (2 × 10k sequential runs)
Batch size
16
Learning rate
5e-4
LoRA rank
32
LoRA dropout
0.0
Quantization
false
Image augmentation
true
Shuffle buffer
10,000
Hardware
NVIDIA GH200 (ARM64/aarch64, 480GB HBM3)
Resuming from a checkpoint
finetune.py has no native resume support. We resumed by setting VLA_PATH to the merged run 2 checkpoint. The model starts from those weights (all learned knowledge preserved); only the step counter and optimizer state reset. This is equivalent to a warm-start.
Key Fix: Gradient Clipping
The upstream vla-scripts/finetune.py in the OpenVLA repo is missing gradient norm clipping before optimizer.step(). A single batch with an outlier action can produce an unclamped gradient large enough to permanently corrupt the model weights.
What we saw: In run 1 (no clipping), loss jumped from 2.2 → 5.2 at step ~6k and action accuracy collapsed from 0.35 → 0.0 and never recovered. Final score: 39%.
The fix:
python
1# Add before optimizer.step() in the training loop:2torch.nn.utils.clip_grad_norm_(trainable_params, max_norm=1.0)3optimizer.step()
With clipping, both runs (2 and 3) converged cleanly. See openvla/openvla#333 for the full analysis with loss curves.
Usage
This is a merged checkpoint — LoRA adapter weights are already merged into the base model. No PEFT library needed.
The unnorm_key="libero_spatial" is required — it selects the action normalization statistics from dataset_statistics.json that were computed from the LIBERO-Spatial training data.
